Unsupervised Person Search in Surveillance Videos by Hybrid Contrastive Learning

With the proliferation of surveillance videos and massive amount of data, manual analysis of surveillance videos becomes increasingly impractical. Since no bounding boxes of persons are provided and person identities are not labelled, existing supervised person search methods are inadequate for real-world surveillance scenarios. To address this issue, an unsupervised person search method is proposed by integrating unsupervised learning and hybrid contrastive learning. Firstly, an unsupervised clustering algorithm using graph partition is employed to annotate person identities and generate pseudo labels. Subsequently, a hybrid contrastive learning is proposed to distinguish different persons within a single camera and enhance the similarity of the same person among multiple cameras. Experimental results on a constructed video database demonstrate the search performance of the proposed method and its potential value in practical surveillance applications.